2 * SPDX-License-Identifier: BSD-2-Clause
4 * Copyright (c) 2001 McAfee, Inc.
5 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
8 * This software was developed for the FreeBSD Project by Jonathan Lemon
9 * and McAfee Research, the Security Research Division of McAfee, Inc. under
10 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
11 * DARPA CHATS research program. [2001 McAfee, Inc.]
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 #include <sys/cdefs.h>
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_pcbgroup.h"
41 #include <sys/param.h>
42 #include <sys/systm.h>
44 #include <sys/refcount.h>
45 #include <sys/kernel.h>
46 #include <sys/sysctl.h>
47 #include <sys/limits.h>
49 #include <sys/mutex.h>
50 #include <sys/malloc.h>
52 #include <sys/proc.h> /* for proc0 declaration */
53 #include <sys/random.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/syslog.h>
57 #include <sys/ucred.h>
60 #include <crypto/siphash/siphash.h>
65 #include <net/if_var.h>
66 #include <net/route.h>
69 #include <netinet/in.h>
70 #include <netinet/in_kdtrace.h>
71 #include <netinet/in_systm.h>
72 #include <netinet/ip.h>
73 #include <netinet/in_var.h>
74 #include <netinet/in_pcb.h>
75 #include <netinet/ip_var.h>
76 #include <netinet/ip_options.h>
78 #include <netinet/ip6.h>
79 #include <netinet/icmp6.h>
80 #include <netinet6/nd6.h>
81 #include <netinet6/ip6_var.h>
82 #include <netinet6/in6_pcb.h>
84 #include <netinet/tcp.h>
85 #include <netinet/tcp_fastopen.h>
86 #include <netinet/tcp_fsm.h>
87 #include <netinet/tcp_seq.h>
88 #include <netinet/tcp_timer.h>
89 #include <netinet/tcp_var.h>
90 #include <netinet/tcp_syncache.h>
92 #include <netinet6/tcp6_var.h>
95 #include <netinet/toecore.h>
97 #include <netinet/udp.h>
98 #include <netinet/udp_var.h>
100 #include <netipsec/ipsec_support.h>
102 #include <machine/in_cksum.h>
104 #include <security/mac/mac_framework.h>
106 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
107 #define V_tcp_syncookies VNET(tcp_syncookies)
108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
109 &VNET_NAME(tcp_syncookies), 0,
110 "Use TCP SYN cookies if the syncache overflows");
112 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
113 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly)
114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
115 &VNET_NAME(tcp_syncookiesonly), 0,
116 "Use only TCP SYN cookies");
118 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
119 #define V_functions_inherit_listen_socket_stack \
120 VNET(functions_inherit_listen_socket_stack)
121 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
122 CTLFLAG_VNET | CTLFLAG_RW,
123 &VNET_NAME(functions_inherit_listen_socket_stack), 0,
124 "Inherit listen socket's stack");
127 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
130 static void syncache_drop(struct syncache *, struct syncache_head *);
131 static void syncache_free(struct syncache *);
132 static void syncache_insert(struct syncache *, struct syncache_head *);
133 static int syncache_respond(struct syncache *, const struct mbuf *, int);
134 static struct socket *syncache_socket(struct syncache *, struct socket *,
136 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
138 static void syncache_timer(void *);
140 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
141 uint8_t *, uintptr_t);
142 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *);
143 static struct syncache
144 *syncookie_lookup(struct in_conninfo *, struct syncache_head *,
145 struct syncache *, struct tcphdr *, struct tcpopt *,
146 struct socket *, uint16_t);
147 static void syncache_pause(struct in_conninfo *);
148 static void syncache_unpause(void *);
149 static void syncookie_reseed(void *);
151 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
152 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
153 struct socket *lso, uint16_t port);
157 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
158 * 3 retransmits corresponds to a timeout with default values of
159 * tcp_rexmit_initial * ( 1 +
162 * tcp_backoff[3]) + 3 * tcp_rexmit_slop,
163 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms,
164 * the odds are that the user has given up attempting to connect by then.
166 #define SYNCACHE_MAXREXMTS 3
168 /* Arbitrary values */
169 #define TCP_SYNCACHE_HASHSIZE 512
170 #define TCP_SYNCACHE_BUCKETLIMIT 30
172 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
173 #define V_tcp_syncache VNET(tcp_syncache)
175 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
176 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
179 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
180 &VNET_NAME(tcp_syncache.bucket_limit), 0,
181 "Per-bucket hash limit for syncache");
183 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
184 &VNET_NAME(tcp_syncache.cache_limit), 0,
185 "Overall entry limit for syncache");
187 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
188 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
190 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
191 &VNET_NAME(tcp_syncache.hashsize), 0,
192 "Size of TCP syncache hashtable");
195 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
200 new = V_tcp_syncache.rexmt_limit;
201 error = sysctl_handle_int(oidp, &new, 0, req);
202 if ((error == 0) && (req->newptr != NULL)) {
203 if (new > TCP_MAXRXTSHIFT)
206 V_tcp_syncache.rexmt_limit = new;
211 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
212 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
213 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
214 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
215 "Limit on SYN/ACK retransmissions");
217 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
218 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
219 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
220 "Send reset on socket allocation failure");
222 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
224 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
225 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
226 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
229 * Requires the syncache entry to be already removed from the bucket list.
232 syncache_free(struct syncache *sc)
236 (void) m_free(sc->sc_ipopts);
240 mac_syncache_destroy(&sc->sc_label);
243 uma_zfree(V_tcp_syncache.zone, sc);
251 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
252 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
253 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
254 V_tcp_syncache.hash_secret = arc4random();
256 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
257 &V_tcp_syncache.hashsize);
258 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
259 &V_tcp_syncache.bucket_limit);
260 if (!powerof2(V_tcp_syncache.hashsize) ||
261 V_tcp_syncache.hashsize == 0) {
262 printf("WARNING: syncache hash size is not a power of 2.\n");
263 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
265 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
268 V_tcp_syncache.cache_limit =
269 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
270 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
271 &V_tcp_syncache.cache_limit);
273 /* Allocate the hash table. */
274 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
275 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
278 V_tcp_syncache.vnet = curvnet;
281 /* Initialize the hash buckets. */
282 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
283 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
284 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
286 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
287 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
288 V_tcp_syncache.hashbase[i].sch_length = 0;
289 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
290 V_tcp_syncache.hashbase[i].sch_last_overflow =
291 -(SYNCOOKIE_LIFETIME + 1);
294 /* Create the syncache entry zone. */
295 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
296 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
297 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
298 V_tcp_syncache.cache_limit);
300 /* Start the SYN cookie reseeder callout. */
301 callout_init(&V_tcp_syncache.secret.reseed, 1);
302 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
303 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
304 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
305 syncookie_reseed, &V_tcp_syncache);
307 /* Initialize the pause machinery. */
308 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
309 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
311 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
312 V_tcp_syncache.pause_backoff = 0;
313 V_tcp_syncache.paused = false;
318 syncache_destroy(void)
320 struct syncache_head *sch;
321 struct syncache *sc, *nsc;
325 * Stop the re-seed timer before freeing resources. No need to
326 * possibly schedule it another time.
328 callout_drain(&V_tcp_syncache.secret.reseed);
330 /* Stop the SYN cache pause callout. */
331 mtx_lock(&V_tcp_syncache.pause_mtx);
332 if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
333 mtx_unlock(&V_tcp_syncache.pause_mtx);
334 callout_drain(&V_tcp_syncache.pause_co);
336 mtx_unlock(&V_tcp_syncache.pause_mtx);
338 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
339 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
340 sch = &V_tcp_syncache.hashbase[i];
341 callout_drain(&sch->sch_timer);
344 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
345 syncache_drop(sc, sch);
347 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
348 ("%s: sch->sch_bucket not empty", __func__));
349 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
350 __func__, sch->sch_length));
351 mtx_destroy(&sch->sch_mtx);
354 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
355 ("%s: cache_count not 0", __func__));
357 /* Free the allocated global resources. */
358 uma_zdestroy(V_tcp_syncache.zone);
359 free(V_tcp_syncache.hashbase, M_SYNCACHE);
360 mtx_destroy(&V_tcp_syncache.pause_mtx);
365 * Inserts a syncache entry into the specified bucket row.
366 * Locks and unlocks the syncache_head autonomously.
369 syncache_insert(struct syncache *sc, struct syncache_head *sch)
371 struct syncache *sc2;
376 * Make sure that we don't overflow the per-bucket limit.
377 * If the bucket is full, toss the oldest element.
379 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
380 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
381 ("sch->sch_length incorrect"));
382 syncache_pause(&sc->sc_inc);
383 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
384 sch->sch_last_overflow = time_uptime;
385 syncache_drop(sc2, sch);
388 /* Put it into the bucket. */
389 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
393 if (ADDED_BY_TOE(sc)) {
394 struct toedev *tod = sc->sc_tod;
396 tod->tod_syncache_added(tod, sc->sc_todctx);
400 /* Reinitialize the bucket row's timer. */
401 if (sch->sch_length == 1)
402 sch->sch_nextc = ticks + INT_MAX;
403 syncache_timeout(sc, sch, 1);
407 TCPSTATES_INC(TCPS_SYN_RECEIVED);
408 TCPSTAT_INC(tcps_sc_added);
412 * Remove and free entry from syncache bucket row.
413 * Expects locked syncache head.
416 syncache_drop(struct syncache *sc, struct syncache_head *sch)
419 SCH_LOCK_ASSERT(sch);
421 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
422 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
426 if (ADDED_BY_TOE(sc)) {
427 struct toedev *tod = sc->sc_tod;
429 tod->tod_syncache_removed(tod, sc->sc_todctx);
437 * Engage/reengage time on bucket row.
440 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
444 if (sc->sc_rxmits == 0)
445 rexmt = tcp_rexmit_initial;
448 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
449 tcp_rexmit_min, TCPTV_REXMTMAX);
450 sc->sc_rxttime = ticks + rexmt;
452 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
453 sch->sch_nextc = sc->sc_rxttime;
455 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
456 syncache_timer, (void *)sch);
461 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
462 * If we have retransmitted an entry the maximum number of times, expire it.
463 * One separate timer for each bucket row.
466 syncache_timer(void *xsch)
468 struct syncache_head *sch = (struct syncache_head *)xsch;
469 struct syncache *sc, *nsc;
470 struct epoch_tracker et;
475 CURVNET_SET(sch->sch_sc->vnet);
477 /* NB: syncache_head has already been locked by the callout. */
478 SCH_LOCK_ASSERT(sch);
481 * In the following cycle we may remove some entries and/or
482 * advance some timeouts, so re-initialize the bucket timer.
484 sch->sch_nextc = tick + INT_MAX;
487 * If we have paused processing, unconditionally remove
488 * all syncache entries.
490 mtx_lock(&V_tcp_syncache.pause_mtx);
491 paused = V_tcp_syncache.paused;
492 mtx_unlock(&V_tcp_syncache.pause_mtx);
494 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
496 syncache_drop(sc, sch);
500 * We do not check if the listen socket still exists
501 * and accept the case where the listen socket may be
502 * gone by the time we resend the SYN/ACK. We do
503 * not expect this to happens often. If it does,
504 * then the RST will be sent by the time the remote
505 * host does the SYN/ACK->ACK.
507 if (TSTMP_GT(sc->sc_rxttime, tick)) {
508 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
509 sch->sch_nextc = sc->sc_rxttime;
512 if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
513 sc->sc_flags &= ~SCF_ECN;
515 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
516 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
517 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
518 "giving up and removing syncache entry\n",
522 syncache_drop(sc, sch);
523 TCPSTAT_INC(tcps_sc_stale);
526 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
527 log(LOG_DEBUG, "%s; %s: Response timeout, "
528 "retransmitting (%u) SYN|ACK\n",
529 s, __func__, sc->sc_rxmits);
534 syncache_respond(sc, NULL, TH_SYN|TH_ACK);
536 TCPSTAT_INC(tcps_sc_retransmitted);
537 syncache_timeout(sc, sch, 0);
539 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
540 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
541 syncache_timer, (void *)(sch));
546 * Returns true if the system is only using cookies at the moment.
547 * This could be due to a sysadmin decision to only use cookies, or it
548 * could be due to the system detecting an attack.
551 syncache_cookiesonly(void)
554 return (V_tcp_syncookies && (V_tcp_syncache.paused ||
555 V_tcp_syncookiesonly));
559 * Find the hash bucket for the given connection.
561 static struct syncache_head *
562 syncache_hashbucket(struct in_conninfo *inc)
567 * The hash is built on foreign port + local port + foreign address.
568 * We rely on the fact that struct in_conninfo starts with 16 bits
569 * of foreign port, then 16 bits of local port then followed by 128
570 * bits of foreign address. In case of IPv4 address, the first 3
571 * 32-bit words of the address always are zeroes.
573 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
574 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
576 return (&V_tcp_syncache.hashbase[hash]);
580 * Find an entry in the syncache.
581 * Returns always with locked syncache_head plus a matching entry or NULL.
583 static struct syncache *
584 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
587 struct syncache_head *sch;
589 *schp = sch = syncache_hashbucket(inc);
592 /* Circle through bucket row to find matching entry. */
593 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
594 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
595 sizeof(struct in_endpoints)) == 0)
598 return (sc); /* Always returns with locked sch. */
602 * This function is called when we get a RST for a
603 * non-existent connection, so that we can see if the
604 * connection is in the syn cache. If it is, zap it.
605 * If required send a challenge ACK.
608 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m,
612 struct syncache_head *sch;
615 if (syncache_cookiesonly())
617 sc = syncache_lookup(inc, &sch); /* returns locked sch */
618 SCH_LOCK_ASSERT(sch);
621 * No corresponding connection was found in syncache.
622 * If syncookies are enabled and possibly exclusively
623 * used, or we are under memory pressure, a valid RST
624 * may not find a syncache entry. In that case we're
625 * done and no SYN|ACK retransmissions will happen.
626 * Otherwise the RST was misdirected or spoofed.
629 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
630 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
631 "syncache entry (possibly syncookie only), "
632 "segment ignored\n", s, __func__);
633 TCPSTAT_INC(tcps_badrst);
637 /* The remote UDP encaps port does not match. */
638 if (sc->sc_port != port) {
639 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
640 log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
641 "syncache entry but non-matching UDP encaps port, "
642 "segment ignored\n", s, __func__);
643 TCPSTAT_INC(tcps_badrst);
648 * If the RST bit is set, check the sequence number to see
649 * if this is a valid reset segment.
652 * In all states except SYN-SENT, all reset (RST) segments
653 * are validated by checking their SEQ-fields. A reset is
654 * valid if its sequence number is in the window.
657 * There are four cases for the acceptability test for an incoming
660 * Segment Receive Test
662 * ------- ------- -------------------------------------------
663 * 0 0 SEG.SEQ = RCV.NXT
664 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
665 * >0 0 not acceptable
666 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
667 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
669 * Note that when receiving a SYN segment in the LISTEN state,
670 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
671 * described in RFC 793, page 66.
673 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
674 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
675 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
676 if (V_tcp_insecure_rst ||
677 th->th_seq == sc->sc_irs + 1) {
678 syncache_drop(sc, sch);
679 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
681 "%s; %s: Our SYN|ACK was rejected, "
682 "connection attempt aborted by remote "
685 TCPSTAT_INC(tcps_sc_reset);
687 TCPSTAT_INC(tcps_badrst);
688 /* Send challenge ACK. */
689 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
690 log(LOG_DEBUG, "%s; %s: RST with invalid "
691 " SEQ %u != NXT %u (+WND %u), "
692 "sending challenge ACK\n",
694 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
695 syncache_respond(sc, m, TH_ACK);
698 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
699 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
700 "NXT %u (+WND %u), segment ignored\n",
702 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
703 TCPSTAT_INC(tcps_badrst);
713 syncache_badack(struct in_conninfo *inc, uint16_t port)
716 struct syncache_head *sch;
718 if (syncache_cookiesonly())
720 sc = syncache_lookup(inc, &sch); /* returns locked sch */
721 SCH_LOCK_ASSERT(sch);
722 if ((sc != NULL) && (sc->sc_port == port)) {
723 syncache_drop(sc, sch);
724 TCPSTAT_INC(tcps_sc_badack);
730 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
733 struct syncache_head *sch;
735 if (syncache_cookiesonly())
737 sc = syncache_lookup(inc, &sch); /* returns locked sch */
738 SCH_LOCK_ASSERT(sch);
742 /* If the port != sc_port, then it's a bogus ICMP msg */
743 if (port != sc->sc_port)
746 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
747 if (ntohl(th_seq) != sc->sc_iss)
751 * If we've rertransmitted 3 times and this is our second error,
752 * we remove the entry. Otherwise, we allow it to continue on.
753 * This prevents us from incorrectly nuking an entry during a
754 * spurious network outage.
758 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
759 sc->sc_flags |= SCF_UNREACH;
762 syncache_drop(sc, sch);
763 TCPSTAT_INC(tcps_sc_unreach);
769 * Build a new TCP socket structure from a syncache entry.
771 * On success return the newly created socket with its underlying inp locked.
773 static struct socket *
774 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
776 struct tcp_function_block *blk;
777 struct inpcb *inp = NULL;
786 * Ok, create the full blown connection, and set things up
787 * as they would have been set up if we had created the
788 * connection when the SYN arrived. If we can't create
789 * the connection, abort it.
791 so = sonewconn(lso, 0);
794 * Drop the connection; we will either send a RST or
795 * have the peer retransmit its SYN again after its
798 TCPSTAT_INC(tcps_listendrop);
799 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
800 log(LOG_DEBUG, "%s; %s: Socket create failed "
801 "due to limits or memory shortage\n",
808 mac_socketpeer_set_from_mbuf(m, so);
812 inp->inp_inc.inc_fibnum = so->so_fibnum;
815 * Exclusive pcbinfo lock is not required in syncache socket case even
816 * if two inpcb locks can be acquired simultaneously:
817 * - the inpcb in LISTEN state,
818 * - the newly created inp.
820 * In this case, an inp cannot be at same time in LISTEN state and
821 * just created by an accept() call.
823 INP_HASH_WLOCK(&V_tcbinfo);
825 /* Insert new socket into PCB hash list. */
826 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
828 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
829 inp->inp_vflag &= ~INP_IPV4;
830 inp->inp_vflag |= INP_IPV6;
831 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
833 inp->inp_vflag &= ~INP_IPV6;
834 inp->inp_vflag |= INP_IPV4;
836 inp->inp_ip_ttl = sc->sc_ip_ttl;
837 inp->inp_ip_tos = sc->sc_ip_tos;
838 inp->inp_laddr = sc->sc_inc.inc_laddr;
844 * If there's an mbuf and it has a flowid, then let's initialise the
845 * inp with that particular flowid.
847 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
848 inp->inp_flowid = m->m_pkthdr.flowid;
849 inp->inp_flowtype = M_HASHTYPE_GET(m);
851 inp->inp_numa_domain = m->m_pkthdr.numa_domain;
855 inp->inp_lport = sc->sc_inc.inc_lport;
857 if (inp->inp_vflag & INP_IPV6PROTO) {
858 struct inpcb *oinp = sotoinpcb(lso);
861 * Inherit socket options from the listening socket.
862 * Note that in6p_inputopts are not (and should not be)
863 * copied, since it stores previously received options and is
864 * used to detect if each new option is different than the
865 * previous one and hence should be passed to a user.
866 * If we copied in6p_inputopts, a user would not be able to
867 * receive options just after calling the accept system call.
869 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
870 if (oinp->in6p_outputopts)
871 inp->in6p_outputopts =
872 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
873 inp->in6p_hops = oinp->in6p_hops;
876 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
877 struct in6_addr laddr6;
878 struct sockaddr_in6 sin6;
880 sin6.sin6_family = AF_INET6;
881 sin6.sin6_len = sizeof(sin6);
882 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
883 sin6.sin6_port = sc->sc_inc.inc_fport;
884 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
885 laddr6 = inp->in6p_laddr;
886 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
887 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
888 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
889 thread0.td_ucred, m, false)) != 0) {
890 inp->in6p_laddr = laddr6;
891 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
892 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
897 INP_HASH_WUNLOCK(&V_tcbinfo);
900 /* Override flowlabel from in6_pcbconnect. */
901 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
902 inp->inp_flow |= sc->sc_flowlabel;
905 #if defined(INET) && defined(INET6)
910 struct in_addr laddr;
911 struct sockaddr_in sin;
913 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
915 if (inp->inp_options == NULL) {
916 inp->inp_options = sc->sc_ipopts;
917 sc->sc_ipopts = NULL;
920 sin.sin_family = AF_INET;
921 sin.sin_len = sizeof(sin);
922 sin.sin_addr = sc->sc_inc.inc_faddr;
923 sin.sin_port = sc->sc_inc.inc_fport;
924 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
925 laddr = inp->inp_laddr;
926 if (inp->inp_laddr.s_addr == INADDR_ANY)
927 inp->inp_laddr = sc->sc_inc.inc_laddr;
928 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
929 thread0.td_ucred, m, false)) != 0) {
930 inp->inp_laddr = laddr;
931 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
932 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
937 INP_HASH_WUNLOCK(&V_tcbinfo);
942 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
943 /* Copy old policy into new socket's. */
944 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
945 printf("syncache_socket: could not copy policy\n");
947 INP_HASH_WUNLOCK(&V_tcbinfo);
949 tcp_state_change(tp, TCPS_SYN_RECEIVED);
950 tp->iss = sc->sc_iss;
951 tp->irs = sc->sc_irs;
952 tp->t_port = sc->sc_port;
955 blk = sototcpcb(lso)->t_fb;
956 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
958 * Our parents t_fb was not the default,
959 * we need to release our ref on tp->t_fb and
960 * pickup one on the new entry.
962 struct tcp_function_block *rblk;
964 rblk = find_and_ref_tcp_fb(blk);
965 KASSERT(rblk != NULL,
966 ("cannot find blk %p out of syncache?", blk));
967 if (tp->t_fb->tfb_tcp_fb_fini)
968 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
969 refcount_release(&tp->t_fb->tfb_refcnt);
972 * XXXrrs this is quite dangerous, it is possible
973 * for the new function to fail to init. We also
974 * are not asking if the handoff_is_ok though at
975 * the very start thats probalbly ok.
977 if (tp->t_fb->tfb_tcp_fb_init) {
978 (*tp->t_fb->tfb_tcp_fb_init)(tp);
981 tp->snd_wl1 = sc->sc_irs;
982 tp->snd_max = tp->iss + 1;
983 tp->snd_nxt = tp->iss + 1;
984 tp->rcv_up = sc->sc_irs + 1;
985 tp->rcv_wnd = sc->sc_wnd;
986 tp->rcv_adv += tp->rcv_wnd;
987 tp->last_ack_sent = tp->rcv_nxt;
989 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
990 if (sc->sc_flags & SCF_NOOPT)
991 tp->t_flags |= TF_NOOPT;
993 if (sc->sc_flags & SCF_WINSCALE) {
994 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
995 tp->snd_scale = sc->sc_requested_s_scale;
996 tp->request_r_scale = sc->sc_requested_r_scale;
998 if (sc->sc_flags & SCF_TIMESTAMP) {
999 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
1000 tp->ts_recent = sc->sc_tsreflect;
1001 tp->ts_recent_age = tcp_ts_getticks();
1002 tp->ts_offset = sc->sc_tsoff;
1004 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1005 if (sc->sc_flags & SCF_SIGNATURE)
1006 tp->t_flags |= TF_SIGNATURE;
1008 if (sc->sc_flags & SCF_SACK)
1009 tp->t_flags |= TF_SACK_PERMIT;
1012 if (sc->sc_flags & SCF_ECN)
1013 tp->t_flags2 |= TF2_ECN_PERMIT;
1016 * Set up MSS and get cached values from tcp_hostcache.
1017 * This might overwrite some of the defaults we just set.
1019 tcp_mss(tp, sc->sc_peer_mss);
1022 * If the SYN,ACK was retransmitted, indicate that CWND to be
1023 * limited to one segment in cc_conn_init().
1024 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
1026 if (sc->sc_rxmits > 1)
1031 * Allow a TOE driver to install its hooks. Note that we hold the
1032 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
1033 * new connection before the TOE driver has done its thing.
1035 if (ADDED_BY_TOE(sc)) {
1036 struct toedev *tod = sc->sc_tod;
1038 tod->tod_offload_socket(tod, sc->sc_todctx, so);
1042 * Copy and activate timers.
1044 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1045 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1046 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1047 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1048 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1050 TCPSTAT_INC(tcps_accepts);
1062 * This function gets called when we receive an ACK for a
1063 * socket in the LISTEN state. We look up the connection
1064 * in the syncache, and if its there, we pull it out of
1065 * the cache and turn it into a full-blown connection in
1066 * the SYN-RECEIVED state.
1068 * On syncache_socket() success the newly created socket
1069 * has its underlying inp locked.
1072 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1073 struct socket **lsop, struct mbuf *m, uint16_t port)
1075 struct syncache *sc;
1076 struct syncache_head *sch;
1077 struct syncache scs;
1082 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1083 ("%s: can handle only ACK", __func__));
1085 if (syncache_cookiesonly()) {
1087 sch = syncache_hashbucket(inc);
1090 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1092 SCH_LOCK_ASSERT(sch);
1097 * Test code for syncookies comparing the syncache stored
1098 * values with the reconstructed values from the cookie.
1101 syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1106 * There is no syncache entry, so see if this ACK is
1107 * a returning syncookie. To do this, first:
1108 * A. Check if syncookies are used in case of syncache
1110 * B. See if this socket has had a syncache entry dropped in
1111 * the recent past. We don't want to accept a bogus
1112 * syncookie if we've never received a SYN or accept it
1114 * C. check that the syncookie is valid. If it is, then
1115 * cobble up a fake syncache entry, and return.
1117 if (locked && !V_tcp_syncookies) {
1119 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1120 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1121 "segment rejected (syncookies disabled)\n",
1125 if (locked && !V_tcp_syncookiesonly &&
1126 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1128 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1129 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1130 "segment rejected (no syncache entry)\n",
1134 bzero(&scs, sizeof(scs));
1135 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port);
1139 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1140 log(LOG_DEBUG, "%s; %s: Segment failed "
1141 "SYNCOOKIE authentication, segment rejected "
1142 "(probably spoofed)\n", s, __func__);
1145 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1146 /* If received ACK has MD5 signature, check it. */
1147 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1148 (!TCPMD5_ENABLED() ||
1149 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1151 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1152 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1153 "MD5 signature doesn't match.\n",
1157 TCPSTAT_INC(tcps_sig_err_sigopt);
1158 return (-1); /* Do not send RST */
1160 #endif /* TCP_SIGNATURE */
1162 if (sc->sc_port != port) {
1166 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1168 * If listening socket requested TCP digests, check that
1169 * received ACK has signature and it is correct.
1170 * If not, drop the ACK and leave sc entry in th cache,
1171 * because SYN was received with correct signature.
1173 if (sc->sc_flags & SCF_SIGNATURE) {
1174 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1176 TCPSTAT_INC(tcps_sig_err_nosigopt);
1178 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1179 log(LOG_DEBUG, "%s; %s: Segment "
1180 "rejected, MD5 signature wasn't "
1181 "provided.\n", s, __func__);
1184 return (-1); /* Do not send RST */
1186 if (!TCPMD5_ENABLED() ||
1187 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1188 /* Doesn't match or no SA */
1190 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1191 log(LOG_DEBUG, "%s; %s: Segment "
1192 "rejected, MD5 signature doesn't "
1193 "match.\n", s, __func__);
1196 return (-1); /* Do not send RST */
1199 #endif /* TCP_SIGNATURE */
1202 * RFC 7323 PAWS: If we have a timestamp on this segment and
1203 * it's less than ts_recent, drop it.
1204 * XXXMT: RFC 7323 also requires to send an ACK.
1205 * In tcp_input.c this is only done for TCP segments
1206 * with user data, so be consistent here and just drop
1209 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1210 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1212 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1214 "%s; %s: SEG.TSval %u < TS.Recent %u, "
1215 "segment dropped\n", s, __func__,
1216 to->to_tsval, sc->sc_tsreflect);
1219 return (-1); /* Do not send RST */
1223 * If timestamps were not negotiated during SYN/ACK and a
1224 * segment with a timestamp is received, ignore the
1225 * timestamp and process the packet normally.
1226 * See section 3.2 of RFC 7323.
1228 if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1229 (to->to_flags & TOF_TS)) {
1230 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1231 log(LOG_DEBUG, "%s; %s: Timestamp not "
1232 "expected, segment processed normally\n",
1240 * If timestamps were negotiated during SYN/ACK and a
1241 * segment without a timestamp is received, silently drop
1242 * the segment, unless the missing timestamps are tolerated.
1243 * See section 3.2 of RFC 7323.
1245 if ((sc->sc_flags & SCF_TIMESTAMP) &&
1246 !(to->to_flags & TOF_TS)) {
1247 if (V_tcp_tolerate_missing_ts) {
1248 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1250 "%s; %s: Timestamp missing, "
1251 "segment processed normally\n",
1257 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1259 "%s; %s: Timestamp missing, "
1260 "segment silently dropped\n",
1264 return (-1); /* Do not send RST */
1269 * Pull out the entry to unlock the bucket row.
1271 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1272 * tcp_state_change(). The tcpcb is not existent at this
1273 * moment. A new one will be allocated via syncache_socket->
1274 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1275 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1277 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1278 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1281 if (ADDED_BY_TOE(sc)) {
1282 struct toedev *tod = sc->sc_tod;
1284 tod->tod_syncache_removed(tod, sc->sc_todctx);
1291 * Segment validation:
1292 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1294 if (th->th_ack != sc->sc_iss + 1) {
1295 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1296 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1297 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1302 * The SEQ must fall in the window starting at the received
1303 * initial receive sequence number + 1 (the SYN).
1305 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1306 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1307 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1308 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1309 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1313 *lsop = syncache_socket(sc, *lsop, m);
1316 TCPSTAT_INC(tcps_sc_aborted);
1318 TCPSTAT_INC(tcps_sc_completed);
1320 /* how do we find the inp for the new socket? */
1325 if (sc != NULL && sc != &scs)
1334 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1335 uint64_t response_cookie)
1339 unsigned int *pending_counter;
1343 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1344 *lsop = syncache_socket(sc, *lsop, m);
1345 if (*lsop == NULL) {
1346 TCPSTAT_INC(tcps_sc_aborted);
1347 atomic_subtract_int(pending_counter, 1);
1349 soisconnected(*lsop);
1350 inp = sotoinpcb(*lsop);
1351 tp = intotcpcb(inp);
1352 tp->t_flags |= TF_FASTOPEN;
1353 tp->t_tfo_cookie.server = response_cookie;
1354 tp->snd_max = tp->iss;
1355 tp->snd_nxt = tp->iss;
1356 tp->t_tfo_pending = pending_counter;
1357 TCPSTATES_INC(TCPS_SYN_RECEIVED);
1358 TCPSTAT_INC(tcps_sc_completed);
1363 * Given a LISTEN socket and an inbound SYN request, add
1364 * this to the syn cache, and send back a segment:
1365 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1368 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1369 * Doing so would require that we hold onto the data and deliver it
1370 * to the application. However, if we are the target of a SYN-flood
1371 * DoS attack, an attacker could send data which would eventually
1372 * consume all available buffer space if it were ACKed. By not ACKing
1373 * the data, we avoid this DoS scenario.
1375 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1376 * cookie is processed and a new socket is created. In this case, any data
1377 * accompanying the SYN will be queued to the socket by tcp_input() and will
1378 * be ACKed either when the application sends response data or the delayed
1379 * ACK timer expires, whichever comes first.
1382 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1383 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1384 void *todctx, uint8_t iptos, uint16_t port)
1388 struct syncache *sc = NULL;
1389 struct syncache_head *sch;
1390 struct mbuf *ipopts = NULL;
1392 int win, ip_ttl, ip_tos;
1396 int autoflowlabel = 0;
1399 struct label *maclabel;
1401 struct syncache scs;
1403 uint64_t tfo_response_cookie;
1404 unsigned int *tfo_pending = NULL;
1405 int tfo_cookie_valid = 0;
1406 int tfo_response_cookie_valid = 0;
1409 INP_WLOCK_ASSERT(inp); /* listen socket */
1410 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1411 ("%s: unexpected tcp flags", __func__));
1414 * Combine all so/tp operations very early to drop the INP lock as
1418 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1420 cred = crhold(so->so_cred);
1423 if (inc->inc_flags & INC_ISIPV6) {
1424 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1427 ip_ttl = in6_selecthlim(inp, NULL);
1428 if ((inp->in6p_outputopts == NULL) ||
1429 (inp->in6p_outputopts->ip6po_tclass == -1)) {
1432 ip_tos = inp->in6p_outputopts->ip6po_tclass;
1436 #if defined(INET6) && defined(INET)
1441 ip_ttl = inp->inp_ip_ttl;
1442 ip_tos = inp->inp_ip_tos;
1445 win = so->sol_sbrcv_hiwat;
1446 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1448 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1449 (tp->t_tfo_pending != NULL) &&
1450 (to->to_flags & TOF_FASTOPEN)) {
1452 * Limit the number of pending TFO connections to
1453 * approximately half of the queue limit. This prevents TFO
1454 * SYN floods from starving the service by filling the
1455 * listen queue with bogus TFO connections.
1457 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1458 (so->sol_qlimit / 2)) {
1461 result = tcp_fastopen_check_cookie(inc,
1462 to->to_tfo_cookie, to->to_tfo_len,
1463 &tfo_response_cookie);
1464 tfo_cookie_valid = (result > 0);
1465 tfo_response_cookie_valid = (result >= 0);
1469 * Remember the TFO pending counter as it will have to be
1470 * decremented below if we don't make it to syncache_tfo_expand().
1472 tfo_pending = tp->t_tfo_pending;
1475 /* By the time we drop the lock these should no longer be used. */
1480 if (mac_syncache_init(&maclabel) != 0) {
1484 mac_syncache_create(maclabel, inp);
1486 if (!tfo_cookie_valid)
1490 * Remember the IP options, if any.
1493 if (!(inc->inc_flags & INC_ISIPV6))
1496 ipopts = (m) ? ip_srcroute(m) : NULL;
1501 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1503 * When the socket is TCP-MD5 enabled check that,
1504 * - a signed packet is valid
1505 * - a non-signed packet does not have a security association
1507 * If a signed packet fails validation or a non-signed packet has a
1508 * security association, the packet will be dropped.
1510 if (ltflags & TF_SIGNATURE) {
1511 if (to->to_flags & TOF_SIGNATURE) {
1512 if (!TCPMD5_ENABLED() ||
1513 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1516 if (TCPMD5_ENABLED() &&
1517 TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1520 } else if (to->to_flags & TOF_SIGNATURE)
1522 #endif /* TCP_SIGNATURE */
1524 * See if we already have an entry for this connection.
1525 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1527 * XXX: should the syncache be re-initialized with the contents
1528 * of the new SYN here (which may have different options?)
1530 * XXX: We do not check the sequence number to see if this is a
1531 * real retransmit or a new connection attempt. The question is
1532 * how to handle such a case; either ignore it as spoofed, or
1533 * drop the current entry and create a new one?
1535 if (syncache_cookiesonly()) {
1537 sch = syncache_hashbucket(inc);
1540 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1542 SCH_LOCK_ASSERT(sch);
1545 if (tfo_cookie_valid)
1547 TCPSTAT_INC(tcps_sc_dupsyn);
1550 * If we were remembering a previous source route,
1551 * forget it and use the new one we've been given.
1554 (void) m_free(sc->sc_ipopts);
1555 sc->sc_ipopts = ipopts;
1558 * Update timestamp if present.
1560 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1561 sc->sc_tsreflect = to->to_tsval;
1563 sc->sc_flags &= ~SCF_TIMESTAMP;
1565 * Disable ECN if needed.
1567 if ((sc->sc_flags & SCF_ECN) &&
1568 ((th->th_flags & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) {
1569 sc->sc_flags &= ~SCF_ECN;
1573 * Since we have already unconditionally allocated label
1574 * storage, free it up. The syncache entry will already
1575 * have an initialized label we can use.
1577 mac_syncache_destroy(&maclabel);
1579 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1580 /* Retransmit SYN|ACK and reset retransmit count. */
1581 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1582 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1583 "resetting timer and retransmitting SYN|ACK\n",
1587 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1589 syncache_timeout(sc, sch, 1);
1590 TCPSTAT_INC(tcps_sndacks);
1591 TCPSTAT_INC(tcps_sndtotal);
1597 if (tfo_cookie_valid) {
1598 bzero(&scs, sizeof(scs));
1604 * Skip allocating a syncache entry if we are just going to discard
1608 bzero(&scs, sizeof(scs));
1611 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1614 * The zone allocator couldn't provide more entries.
1615 * Treat this as if the cache was full; drop the oldest
1616 * entry and insert the new one.
1618 TCPSTAT_INC(tcps_sc_zonefail);
1619 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1620 sch->sch_last_overflow = time_uptime;
1621 syncache_drop(sc, sch);
1622 syncache_pause(inc);
1624 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1626 if (V_tcp_syncookies) {
1627 bzero(&scs, sizeof(scs));
1631 ("%s: bucket unexpectedly unlocked",
1635 (void) m_free(ipopts);
1642 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1643 sc->sc_tfo_cookie = &tfo_response_cookie;
1646 * Fill in the syncache values.
1649 sc->sc_label = maclabel;
1654 sc->sc_ipopts = ipopts;
1655 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1656 sc->sc_ip_tos = ip_tos;
1657 sc->sc_ip_ttl = ip_ttl;
1660 sc->sc_todctx = todctx;
1662 sc->sc_irs = th->th_seq;
1664 sc->sc_flowlabel = 0;
1667 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1668 * win was derived from socket earlier in the function.
1671 win = imin(win, TCP_MAXWIN);
1674 if (V_tcp_do_rfc1323 &&
1675 !(ltflags & TF_NOOPT)) {
1677 * A timestamp received in a SYN makes
1678 * it ok to send timestamp requests and replies.
1680 if (to->to_flags & TOF_TS) {
1681 sc->sc_tsreflect = to->to_tsval;
1682 sc->sc_flags |= SCF_TIMESTAMP;
1683 sc->sc_tsoff = tcp_new_ts_offset(inc);
1685 if (to->to_flags & TOF_SCALE) {
1689 * Pick the smallest possible scaling factor that
1690 * will still allow us to scale up to sb_max, aka
1691 * kern.ipc.maxsockbuf.
1693 * We do this because there are broken firewalls that
1694 * will corrupt the window scale option, leading to
1695 * the other endpoint believing that our advertised
1696 * window is unscaled. At scale factors larger than
1697 * 5 the unscaled window will drop below 1500 bytes,
1698 * leading to serious problems when traversing these
1701 * With the default maxsockbuf of 256K, a scale factor
1702 * of 3 will be chosen by this algorithm. Those who
1703 * choose a larger maxsockbuf should watch out
1704 * for the compatibility problems mentioned above.
1706 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1707 * or <SYN,ACK>) segment itself is never scaled.
1709 while (wscale < TCP_MAX_WINSHIFT &&
1710 (TCP_MAXWIN << wscale) < sb_max)
1712 sc->sc_requested_r_scale = wscale;
1713 sc->sc_requested_s_scale = to->to_wscale;
1714 sc->sc_flags |= SCF_WINSCALE;
1717 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1719 * If incoming packet has an MD5 signature, flag this in the
1720 * syncache so that syncache_respond() will do the right thing
1723 if (to->to_flags & TOF_SIGNATURE)
1724 sc->sc_flags |= SCF_SIGNATURE;
1725 #endif /* TCP_SIGNATURE */
1726 if (to->to_flags & TOF_SACKPERM)
1727 sc->sc_flags |= SCF_SACK;
1728 if (to->to_flags & TOF_MSS)
1729 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1730 if (ltflags & TF_NOOPT)
1731 sc->sc_flags |= SCF_NOOPT;
1732 if (((th->th_flags & (TH_ECE|TH_CWR)) == (TH_ECE|TH_CWR)) &&
1734 sc->sc_flags |= SCF_ECN;
1736 if (V_tcp_syncookies)
1737 sc->sc_iss = syncookie_generate(sch, sc);
1739 sc->sc_iss = arc4random();
1741 if (autoflowlabel) {
1742 if (V_tcp_syncookies)
1743 sc->sc_flowlabel = sc->sc_iss;
1745 sc->sc_flowlabel = ip6_randomflowlabel();
1746 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1752 if (tfo_cookie_valid) {
1753 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1754 /* INP_WUNLOCK(inp) will be performed by the caller */
1759 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1761 * Do a standard 3-way handshake.
1763 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1764 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1766 else if (sc != &scs)
1767 syncache_insert(sc, sch); /* locks and unlocks sch */
1768 TCPSTAT_INC(tcps_sndacks);
1769 TCPSTAT_INC(tcps_sndtotal);
1773 TCPSTAT_INC(tcps_sc_dropped);
1778 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1785 * If tfo_pending is not NULL here, then a TFO SYN that did not
1786 * result in a new socket was processed and the associated pending
1787 * counter has not yet been decremented. All such TFO processing paths
1788 * transit this point.
1790 if (tfo_pending != NULL)
1791 tcp_fastopen_decrement_counter(tfo_pending);
1798 mac_syncache_destroy(&maclabel);
1804 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment,
1805 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1808 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1810 struct ip *ip = NULL;
1812 struct tcphdr *th = NULL;
1813 struct udphdr *udp = NULL;
1814 int optlen, error = 0; /* Make compiler happy */
1815 u_int16_t hlen, tlen, mssopt, ulen;
1818 struct ip6_hdr *ip6 = NULL;
1825 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1828 tlen = hlen + sizeof(struct tcphdr);
1830 tlen += sizeof(struct udphdr);
1832 /* Determine MSS we advertize to other end of connection. */
1833 mssopt = tcp_mssopt(&sc->sc_inc);
1835 mssopt -= V_tcp_udp_tunneling_overhead;
1836 mssopt = max(mssopt, V_tcp_minmss);
1838 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1839 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1840 ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1841 "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1842 max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1844 /* Create the IP+TCP header from scratch. */
1845 m = m_gethdr(M_NOWAIT, MT_DATA);
1849 mac_syncache_create_mbuf(sc->sc_label, m);
1851 m->m_data += max_linkhdr;
1853 m->m_pkthdr.len = tlen;
1854 m->m_pkthdr.rcvif = NULL;
1857 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1858 ip6 = mtod(m, struct ip6_hdr *);
1859 ip6->ip6_vfc = IPV6_VERSION;
1860 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1861 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1862 ip6->ip6_plen = htons(tlen - hlen);
1863 /* ip6_hlim is set after checksum */
1864 /* Zero out traffic class and flow label. */
1865 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1866 ip6->ip6_flow |= sc->sc_flowlabel;
1867 if (sc->sc_port != 0) {
1868 ip6->ip6_nxt = IPPROTO_UDP;
1869 udp = (struct udphdr *)(ip6 + 1);
1870 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1871 udp->uh_dport = sc->sc_port;
1872 ulen = (tlen - sizeof(struct ip6_hdr));
1873 th = (struct tcphdr *)(udp + 1);
1875 ip6->ip6_nxt = IPPROTO_TCP;
1876 th = (struct tcphdr *)(ip6 + 1);
1878 ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20);
1881 #if defined(INET6) && defined(INET)
1886 ip = mtod(m, struct ip *);
1887 ip->ip_v = IPVERSION;
1888 ip->ip_hl = sizeof(struct ip) >> 2;
1889 ip->ip_len = htons(tlen);
1893 ip->ip_src = sc->sc_inc.inc_laddr;
1894 ip->ip_dst = sc->sc_inc.inc_faddr;
1895 ip->ip_ttl = sc->sc_ip_ttl;
1896 ip->ip_tos = sc->sc_ip_tos;
1899 * See if we should do MTU discovery. Route lookups are
1900 * expensive, so we will only unset the DF bit if:
1902 * 1) path_mtu_discovery is disabled
1903 * 2) the SCF_UNREACH flag has been set
1905 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1906 ip->ip_off |= htons(IP_DF);
1907 if (sc->sc_port == 0) {
1908 ip->ip_p = IPPROTO_TCP;
1909 th = (struct tcphdr *)(ip + 1);
1911 ip->ip_p = IPPROTO_UDP;
1912 udp = (struct udphdr *)(ip + 1);
1913 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1914 udp->uh_dport = sc->sc_port;
1915 ulen = (tlen - sizeof(struct ip));
1916 th = (struct tcphdr *)(udp + 1);
1920 th->th_sport = sc->sc_inc.inc_lport;
1921 th->th_dport = sc->sc_inc.inc_fport;
1924 th->th_seq = htonl(sc->sc_iss);
1926 th->th_seq = htonl(sc->sc_iss + 1);
1927 th->th_ack = htonl(sc->sc_irs + 1);
1928 th->th_off = sizeof(struct tcphdr) >> 2;
1930 th->th_flags = flags;
1931 th->th_win = htons(sc->sc_wnd);
1934 if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) {
1935 th->th_flags |= TH_ECE;
1936 TCPSTAT_INC(tcps_ecn_shs);
1939 /* Tack on the TCP options. */
1940 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1943 if (flags & TH_SYN) {
1945 to.to_flags = TOF_MSS;
1946 if (sc->sc_flags & SCF_WINSCALE) {
1947 to.to_wscale = sc->sc_requested_r_scale;
1948 to.to_flags |= TOF_SCALE;
1950 if (sc->sc_flags & SCF_SACK)
1951 to.to_flags |= TOF_SACKPERM;
1952 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1953 if (sc->sc_flags & SCF_SIGNATURE)
1954 to.to_flags |= TOF_SIGNATURE;
1956 if (sc->sc_tfo_cookie) {
1957 to.to_flags |= TOF_FASTOPEN;
1958 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1959 to.to_tfo_cookie = sc->sc_tfo_cookie;
1960 /* don't send cookie again when retransmitting response */
1961 sc->sc_tfo_cookie = NULL;
1964 if (sc->sc_flags & SCF_TIMESTAMP) {
1965 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1966 to.to_tsecr = sc->sc_tsreflect;
1967 to.to_flags |= TOF_TS;
1969 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1971 /* Adjust headers by option size. */
1972 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1974 m->m_pkthdr.len += optlen;
1976 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1977 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1980 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1981 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1982 if (sc->sc_flags & SCF_SIGNATURE) {
1983 KASSERT(to.to_flags & TOF_SIGNATURE,
1984 ("tcp_addoptions() didn't set tcp_signature"));
1986 /* NOTE: to.to_signature is inside of mbuf */
1987 if (!TCPMD5_ENABLED() ||
1988 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1999 udp->uh_ulen = htons(ulen);
2001 M_SETFIB(m, sc->sc_inc.inc_fibnum);
2003 * If we have peer's SYN and it has a flowid, then let's assign it to
2004 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
2005 * to SYN|ACK due to lack of inp here.
2007 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
2008 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
2009 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
2012 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
2014 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
2015 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2016 udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
2018 th->th_sum = htons(0);
2020 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
2021 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2022 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
2025 ip6->ip6_hlim = sc->sc_ip_ttl;
2027 if (ADDED_BY_TOE(sc)) {
2028 struct toedev *tod = sc->sc_tod;
2030 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2035 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2036 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2039 #if defined(INET6) && defined(INET)
2045 m->m_pkthdr.csum_flags = CSUM_UDP;
2046 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2047 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2048 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2049 th->th_sum = htons(0);
2051 m->m_pkthdr.csum_flags = CSUM_TCP;
2052 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2053 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2054 htons(tlen + optlen - hlen + IPPROTO_TCP));
2057 if (ADDED_BY_TOE(sc)) {
2058 struct toedev *tod = sc->sc_tod;
2060 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2065 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2066 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2073 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2074 * that exceed the capacity of the syncache by avoiding the storage of any
2075 * of the SYNs we receive. Syncookies defend against blind SYN flooding
2076 * attacks where the attacker does not have access to our responses.
2078 * Syncookies encode and include all necessary information about the
2079 * connection setup within the SYN|ACK that we send back. That way we
2080 * can avoid keeping any local state until the ACK to our SYN|ACK returns
2081 * (if ever). Normally the syncache and syncookies are running in parallel
2082 * with the latter taking over when the former is exhausted. When matching
2083 * syncache entry is found the syncookie is ignored.
2085 * The only reliable information persisting the 3WHS is our initial sequence
2086 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
2087 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2088 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
2089 * returns and signifies a legitimate connection if it matches the ACK.
2091 * The available space of 32 bits to store the hash and to encode the SYN
2092 * option information is very tight and we should have at least 24 bits for
2093 * the MAC to keep the number of guesses by blind spoofing reasonably high.
2095 * SYN option information we have to encode to fully restore a connection:
2096 * MSS: is imporant to chose an optimal segment size to avoid IP level
2097 * fragmentation along the path. The common MSS values can be encoded
2098 * in a 3-bit table. Uncommon values are captured by the next lower value
2099 * in the table leading to a slight increase in packetization overhead.
2100 * WSCALE: is necessary to allow large windows to be used for high delay-
2101 * bandwidth product links. Not scaling the window when it was initially
2102 * negotiated is bad for performance as lack of scaling further decreases
2103 * the apparent available send window. We only need to encode the WSCALE
2104 * we received from the remote end. Our end can be recalculated at any
2105 * time. The common WSCALE values can be encoded in a 3-bit table.
2106 * Uncommon values are captured by the next lower value in the table
2107 * making us under-estimate the available window size halving our
2108 * theoretically possible maximum throughput for that connection.
2109 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2110 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2111 * that are included in all segments on a connection. We enable them when
2114 * Security of syncookies and attack vectors:
2116 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2117 * together with the gloabl secret to make it unique per connection attempt.
2118 * Thus any change of any of those parameters results in a different MAC output
2119 * in an unpredictable way unless a collision is encountered. 24 bits of the
2120 * MAC are embedded into the ISS.
2122 * To prevent replay attacks two rotating global secrets are updated with a
2123 * new random value every 15 seconds. The life-time of a syncookie is thus
2126 * Vector 1: Attacking the secret. This requires finding a weakness in the
2127 * MAC itself or the way it is used here. The attacker can do a chosen plain
2128 * text attack by varying and testing the all parameters under his control.
2129 * The strength depends on the size and randomness of the secret, and the
2130 * cryptographic security of the MAC function. Due to the constant updating
2131 * of the secret the attacker has at most 29.999 seconds to find the secret
2132 * and launch spoofed connections. After that he has to start all over again.
2134 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
2135 * size an average of 4,823 attempts are required for a 50% chance of success
2136 * to spoof a single syncookie (birthday collision paradox). However the
2137 * attacker is blind and doesn't know if one of his attempts succeeded unless
2138 * he has a side channel to interfere success from. A single connection setup
2139 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2140 * This many attempts are required for each one blind spoofed connection. For
2141 * every additional spoofed connection he has to launch another N attempts.
2142 * Thus for a sustained rate 100 spoofed connections per second approximately
2143 * 1,800,000 packets per second would have to be sent.
2145 * NB: The MAC function should be fast so that it doesn't become a CPU
2146 * exhaustion attack vector itself.
2149 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2150 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2151 * http://cr.yp.to/syncookies.html (overview)
2152 * http://cr.yp.to/syncookies/archive (details)
2155 * Schematic construction of a syncookie enabled Initial Sequence Number:
2157 * 12345678901234567890123456789012
2158 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2160 * x 24 MAC (truncated)
2161 * W 3 Send Window Scale index
2163 * S 1 SACK permitted
2164 * P 1 Odd/even secret
2168 * Distribution and probability of certain MSS values. Those in between are
2169 * rounded down to the next lower one.
2170 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2171 * .2% .3% 5% 7% 7% 20% 15% 45%
2173 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2176 * Distribution and probability of certain WSCALE values. We have to map the
2177 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2178 * bits based on prevalence of certain values. Where we don't have an exact
2179 * match for are rounded down to the next lower one letting us under-estimate
2180 * the true available window. At the moment this would happen only for the
2181 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2182 * and window size). The absence of the WSCALE option (no scaling in either
2183 * direction) is encoded with index zero.
2184 * [WSCALE values histograms, Allman, 2012]
2185 * X 10 10 35 5 6 14 10% by host
2186 * X 11 4 5 5 18 49 3% by connections
2188 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2191 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
2192 * and good cryptographic properties.
2195 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2196 uint8_t *secbits, uintptr_t secmod)
2199 uint32_t siphash[2];
2201 SipHash24_Init(&ctx);
2202 SipHash_SetKey(&ctx, secbits);
2203 switch (inc->inc_flags & INC_ISIPV6) {
2206 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2207 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2212 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2213 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2217 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2218 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2219 SipHash_Update(&ctx, &irs, sizeof(irs));
2220 SipHash_Update(&ctx, &flags, sizeof(flags));
2221 SipHash_Update(&ctx, &secmod, sizeof(secmod));
2222 SipHash_Final((u_int8_t *)&siphash, &ctx);
2224 return (siphash[0] ^ siphash[1]);
2228 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2230 u_int i, secbit, wscale;
2233 union syncookie cookie;
2237 /* Map our computed MSS into the 3-bit index. */
2238 for (i = nitems(tcp_sc_msstab) - 1;
2239 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2242 cookie.flags.mss_idx = i;
2245 * Map the send window scale into the 3-bit index but only if
2246 * the wscale option was received.
2248 if (sc->sc_flags & SCF_WINSCALE) {
2249 wscale = sc->sc_requested_s_scale;
2250 for (i = nitems(tcp_sc_wstab) - 1;
2251 tcp_sc_wstab[i] > wscale && i > 0;
2254 cookie.flags.wscale_idx = i;
2257 /* Can we do SACK? */
2258 if (sc->sc_flags & SCF_SACK)
2259 cookie.flags.sack_ok = 1;
2261 /* Which of the two secrets to use. */
2262 secbit = V_tcp_syncache.secret.oddeven & 0x1;
2263 cookie.flags.odd_even = secbit;
2265 secbits = V_tcp_syncache.secret.key[secbit];
2266 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2270 * Put the flags into the hash and XOR them to get better ISS number
2271 * variance. This doesn't enhance the cryptographic strength and is
2272 * done to prevent the 8 cookie bits from showing up directly on the
2276 iss |= cookie.cookie ^ (hash >> 24);
2278 TCPSTAT_INC(tcps_sc_sendcookie);
2282 static struct syncache *
2283 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2284 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2285 struct socket *lso, uint16_t port)
2290 int wnd, wscale = 0;
2291 union syncookie cookie;
2294 * Pull information out of SYN-ACK/ACK and revert sequence number
2297 ack = th->th_ack - 1;
2298 seq = th->th_seq - 1;
2301 * Unpack the flags containing enough information to restore the
2304 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2306 /* Which of the two secrets to use. */
2307 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2309 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2311 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2312 if ((ack & ~0xff) != (hash & ~0xff))
2315 /* Fill in the syncache values. */
2317 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2318 sc->sc_ipopts = NULL;
2323 switch (inc->inc_flags & INC_ISIPV6) {
2326 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2327 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2332 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2334 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2339 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2341 /* We can simply recompute receive window scale we sent earlier. */
2342 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2345 /* Only use wscale if it was enabled in the orignal SYN. */
2346 if (cookie.flags.wscale_idx > 0) {
2347 sc->sc_requested_r_scale = wscale;
2348 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2349 sc->sc_flags |= SCF_WINSCALE;
2352 wnd = lso->sol_sbrcv_hiwat;
2354 wnd = imin(wnd, TCP_MAXWIN);
2357 if (cookie.flags.sack_ok)
2358 sc->sc_flags |= SCF_SACK;
2360 if (to->to_flags & TOF_TS) {
2361 sc->sc_flags |= SCF_TIMESTAMP;
2362 sc->sc_tsreflect = to->to_tsval;
2363 sc->sc_tsoff = tcp_new_ts_offset(inc);
2366 if (to->to_flags & TOF_SIGNATURE)
2367 sc->sc_flags |= SCF_SIGNATURE;
2373 TCPSTAT_INC(tcps_sc_recvcookie);
2379 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2380 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2381 struct socket *lso, uint16_t port)
2383 struct syncache scs, *scx;
2386 bzero(&scs, sizeof(scs));
2387 scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port);
2389 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2393 if (sc->sc_peer_mss != scx->sc_peer_mss)
2394 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2395 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2397 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2398 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2399 s, __func__, sc->sc_requested_r_scale,
2400 scx->sc_requested_r_scale);
2402 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2403 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2404 s, __func__, sc->sc_requested_s_scale,
2405 scx->sc_requested_s_scale);
2407 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2408 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2415 #endif /* INVARIANTS */
2418 syncookie_reseed(void *arg)
2420 struct tcp_syncache *sc = arg;
2425 * Reseeding the secret doesn't have to be protected by a lock.
2426 * It only must be ensured that the new random values are visible
2427 * to all CPUs in a SMP environment. The atomic with release
2428 * semantics ensures that.
2430 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2431 secbits = sc->secret.key[secbit];
2432 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2433 atomic_add_rel_int(&sc->secret.oddeven, 1);
2435 /* Reschedule ourself. */
2436 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2440 * We have overflowed a bucket. Let's pause dealing with the syncache.
2441 * This function will increment the bucketoverflow statistics appropriately
2442 * (once per pause when pausing is enabled; otherwise, once per overflow).
2445 syncache_pause(struct in_conninfo *inc)
2451 * 2. Add sysctl read here so we don't get the benefit of this
2452 * change without the new sysctl.
2456 * Try an unlocked read. If we already know that another thread
2457 * has activated the feature, there is no need to proceed.
2459 if (V_tcp_syncache.paused)
2462 /* Are cookied enabled? If not, we can't pause. */
2463 if (!V_tcp_syncookies) {
2464 TCPSTAT_INC(tcps_sc_bucketoverflow);
2469 * We may be the first thread to find an overflow. Get the lock
2470 * and evaluate if we need to take action.
2472 mtx_lock(&V_tcp_syncache.pause_mtx);
2473 if (V_tcp_syncache.paused) {
2474 mtx_unlock(&V_tcp_syncache.pause_mtx);
2478 /* Activate protection. */
2479 V_tcp_syncache.paused = true;
2480 TCPSTAT_INC(tcps_sc_bucketoverflow);
2483 * Determine the last backoff time. If we are seeing a re-newed
2484 * attack within that same time after last reactivating the syncache,
2485 * consider it an extension of the same attack.
2487 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2488 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2489 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2491 V_tcp_syncache.pause_backoff++;
2494 delta = TCP_SYNCACHE_PAUSE_TIME;
2495 V_tcp_syncache.pause_backoff = 0;
2498 /* Log a warning, including IP addresses, if able. */
2500 s = tcp_log_addrs(inc, NULL, NULL, NULL);
2502 s = (const char *)NULL;
2503 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2504 "the next %lld seconds%s%s%s\n", (long long)delta,
2505 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2506 (s != NULL) ? ")" : "");
2507 free(__DECONST(void *, s), M_TCPLOG);
2509 /* Use the calculated delta to set a new pause time. */
2510 V_tcp_syncache.pause_until = time_uptime + delta;
2511 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2513 mtx_unlock(&V_tcp_syncache.pause_mtx);
2516 /* Evaluate whether we need to unpause. */
2518 syncache_unpause(void *arg)
2520 struct tcp_syncache *sc;
2524 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2525 callout_deactivate(&sc->pause_co);
2528 * Check to make sure we are not running early. If the pause
2529 * time has expired, then deactivate the protection.
2531 if ((delta = sc->pause_until - time_uptime) > 0)
2532 callout_schedule(&sc->pause_co, delta * hz);
2538 * Exports the syncache entries to userland so that netstat can display
2539 * them alongside the other sockets. This function is intended to be
2540 * called only from tcp_pcblist.
2542 * Due to concurrency on an active system, the number of pcbs exported
2543 * may have no relation to max_pcbs. max_pcbs merely indicates the
2544 * amount of space the caller allocated for this function to use.
2547 syncache_pcblist(struct sysctl_req *req)
2550 struct syncache *sc;
2551 struct syncache_head *sch;
2554 bzero(&xt, sizeof(xt));
2555 xt.xt_len = sizeof(xt);
2556 xt.t_state = TCPS_SYN_RECEIVED;
2557 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2558 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2559 xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2560 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2562 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2563 sch = &V_tcp_syncache.hashbase[i];
2565 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2566 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2568 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2569 xt.xt_inp.inp_vflag = INP_IPV6;
2571 xt.xt_inp.inp_vflag = INP_IPV4;
2572 xt.xt_encaps_port = sc->sc_port;
2573 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2574 sizeof (struct in_conninfo));
2575 error = SYSCTL_OUT(req, &xt, sizeof xt);